1 //===- InstCombineSimplifyDemanded.cpp ------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file contains logic for simplifying instructions based on information
11 // about how they are used.
13 //===----------------------------------------------------------------------===//
15 #include "InstCombineInternal.h"
16 #include "llvm/Analysis/ValueTracking.h"
17 #include "llvm/IR/IntrinsicInst.h"
18 #include "llvm/IR/PatternMatch.h"
19 #include "llvm/Support/KnownBits.h"
22 using namespace llvm::PatternMatch;
24 #define DEBUG_TYPE "instcombine"
26 /// Check to see if the specified operand of the specified instruction is a
27 /// constant integer. If so, check to see if there are any bits set in the
28 /// constant that are not demanded. If so, shrink the constant and return true.
29 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
30 const APInt &Demanded) {
31 assert(I && "No instruction?");
32 assert(OpNo < I->getNumOperands() && "Operand index too large");
34 // The operand must be a constant integer or splat integer.
35 Value *Op = I->getOperand(OpNo);
37 if (!match(Op, m_APInt(C)))
40 // If there are no bits set that aren't demanded, nothing to do.
41 if (C->isSubsetOf(Demanded))
44 // This instruction is producing bits that are not demanded. Shrink the RHS.
45 I->setOperand(OpNo, ConstantInt::get(Op->getType(), *C & Demanded));
52 /// Inst is an integer instruction that SimplifyDemandedBits knows about. See if
53 /// the instruction has any properties that allow us to simplify its operands.
54 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
55 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
56 KnownBits Known(BitWidth);
57 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
59 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask, Known,
62 if (V == &Inst) return true;
63 replaceInstUsesWith(Inst, V);
67 /// This form of SimplifyDemandedBits simplifies the specified instruction
68 /// operand if possible, updating it in place. It returns true if it made any
69 /// change and false otherwise.
70 bool InstCombiner::SimplifyDemandedBits(Instruction *I, unsigned OpNo,
71 const APInt &DemandedMask,
74 Use &U = I->getOperandUse(OpNo);
75 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask, Known,
77 if (!NewVal) return false;
83 /// This function attempts to replace V with a simpler value based on the
84 /// demanded bits. When this function is called, it is known that only the bits
85 /// set in DemandedMask of the result of V are ever used downstream.
86 /// Consequently, depending on the mask and V, it may be possible to replace V
87 /// with a constant or one of its operands. In such cases, this function does
88 /// the replacement and returns true. In all other cases, it returns false after
89 /// analyzing the expression and setting KnownOne and known to be one in the
90 /// expression. Known.Zero contains all the bits that are known to be zero in
91 /// the expression. These are provided to potentially allow the caller (which
92 /// might recursively be SimplifyDemandedBits itself) to simplify the
94 /// Known.One and Known.Zero always follow the invariant that:
95 /// Known.One & Known.Zero == 0.
96 /// That is, a bit can't be both 1 and 0. Note that the bits in Known.One and
97 /// Known.Zero may only be accurate for those bits set in DemandedMask. Note
98 /// also that the bitwidth of V, DemandedMask, Known.Zero and Known.One must all
101 /// This returns null if it did not change anything and it permits no
102 /// simplification. This returns V itself if it did some simplification of V's
103 /// operands based on the information about what bits are demanded. This returns
104 /// some other non-null value if it found out that V is equal to another value
105 /// in the context where the specified bits are demanded, but not for all users.
106 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
107 KnownBits &Known, unsigned Depth,
109 assert(V != nullptr && "Null pointer of Value???");
110 assert(Depth <= 6 && "Limit Search Depth");
111 uint32_t BitWidth = DemandedMask.getBitWidth();
112 Type *VTy = V->getType();
114 (!VTy->isIntOrIntVectorTy() || VTy->getScalarSizeInBits() == BitWidth) &&
115 Known.getBitWidth() == BitWidth &&
116 "Value *V, DemandedMask and Known must have same BitWidth");
118 if (isa<Constant>(V)) {
119 computeKnownBits(V, Known, Depth, CxtI);
124 if (DemandedMask == 0) // Not demanding any bits from V.
125 return UndefValue::get(VTy);
127 if (Depth == 6) // Limit search depth.
130 Instruction *I = dyn_cast<Instruction>(V);
132 computeKnownBits(V, Known, Depth, CxtI);
133 return nullptr; // Only analyze instructions.
136 // If there are multiple uses of this value and we aren't at the root, then
137 // we can't do any simplifications of the operands, because DemandedMask
138 // only reflects the bits demanded by *one* of the users.
139 if (Depth != 0 && !I->hasOneUse())
140 return SimplifyMultipleUseDemandedBits(I, DemandedMask, Known, Depth, CxtI);
142 KnownBits LHSKnown(BitWidth), RHSKnown(BitWidth);
144 // If this is the root being simplified, allow it to have multiple uses,
145 // just set the DemandedMask to all bits so that we can try to simplify the
146 // operands. This allows visitTruncInst (for example) to simplify the
147 // operand of a trunc without duplicating all the logic below.
148 if (Depth == 0 && !V->hasOneUse())
149 DemandedMask.setAllBits();
151 switch (I->getOpcode()) {
153 computeKnownBits(I, Known, Depth, CxtI);
155 case Instruction::And: {
156 // If either the LHS or the RHS are Zero, the result is zero.
157 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
158 SimplifyDemandedBits(I, 0, DemandedMask & ~RHSKnown.Zero, LHSKnown,
161 assert(!(RHSKnown.Zero & RHSKnown.One) && "Bits known to be one AND zero?");
162 assert(!(LHSKnown.Zero & LHSKnown.One) && "Bits known to be one AND zero?");
164 // Output known-0 are known to be clear if zero in either the LHS | RHS.
165 APInt IKnownZero = RHSKnown.Zero | LHSKnown.Zero;
166 // Output known-1 bits are only known if set in both the LHS & RHS.
167 APInt IKnownOne = RHSKnown.One & LHSKnown.One;
169 // If the client is only demanding bits that we know, return the known
171 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
172 return Constant::getIntegerValue(VTy, IKnownOne);
174 // If all of the demanded bits are known 1 on one side, return the other.
175 // These bits cannot contribute to the result of the 'and'.
176 if (DemandedMask.isSubsetOf(LHSKnown.Zero | RHSKnown.One))
177 return I->getOperand(0);
178 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.One))
179 return I->getOperand(1);
181 // If the RHS is a constant, see if we can simplify it.
182 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnown.Zero))
185 Known.Zero = std::move(IKnownZero);
186 Known.One = std::move(IKnownOne);
189 case Instruction::Or: {
190 // If either the LHS or the RHS are One, the result is One.
191 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
192 SimplifyDemandedBits(I, 0, DemandedMask & ~RHSKnown.One, LHSKnown,
195 assert(!(RHSKnown.Zero & RHSKnown.One) && "Bits known to be one AND zero?");
196 assert(!(LHSKnown.Zero & LHSKnown.One) && "Bits known to be one AND zero?");
198 // Output known-0 bits are only known if clear in both the LHS & RHS.
199 APInt IKnownZero = RHSKnown.Zero & LHSKnown.Zero;
200 // Output known-1 are known. to be set if s.et in either the LHS | RHS.
201 APInt IKnownOne = RHSKnown.One | LHSKnown.One;
203 // If the client is only demanding bits that we know, return the known
205 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
206 return Constant::getIntegerValue(VTy, IKnownOne);
208 // If all of the demanded bits are known zero on one side, return the other.
209 // These bits cannot contribute to the result of the 'or'.
210 if (DemandedMask.isSubsetOf(LHSKnown.One | RHSKnown.Zero))
211 return I->getOperand(0);
212 if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero))
213 return I->getOperand(1);
215 // If the RHS is a constant, see if we can simplify it.
216 if (ShrinkDemandedConstant(I, 1, DemandedMask))
219 Known.Zero = std::move(IKnownZero);
220 Known.One = std::move(IKnownOne);
223 case Instruction::Xor: {
224 if (SimplifyDemandedBits(I, 1, DemandedMask, RHSKnown, Depth + 1) ||
225 SimplifyDemandedBits(I, 0, DemandedMask, LHSKnown, Depth + 1))
227 assert(!(RHSKnown.Zero & RHSKnown.One) && "Bits known to be one AND zero?");
228 assert(!(LHSKnown.Zero & LHSKnown.One) && "Bits known to be one AND zero?");
230 // Output known-0 bits are known if clear or set in both the LHS & RHS.
231 APInt IKnownZero = (RHSKnown.Zero & LHSKnown.Zero) |
232 (RHSKnown.One & LHSKnown.One);
233 // Output known-1 are known to be set if set in only one of the LHS, RHS.
234 APInt IKnownOne = (RHSKnown.Zero & LHSKnown.One) |
235 (RHSKnown.One & LHSKnown.Zero);
237 // If the client is only demanding bits that we know, return the known
239 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
240 return Constant::getIntegerValue(VTy, IKnownOne);
242 // If all of the demanded bits are known zero on one side, return the other.
243 // These bits cannot contribute to the result of the 'xor'.
244 if (DemandedMask.isSubsetOf(RHSKnown.Zero))
245 return I->getOperand(0);
246 if (DemandedMask.isSubsetOf(LHSKnown.Zero))
247 return I->getOperand(1);
249 // If all of the demanded bits are known to be zero on one side or the
250 // other, turn this into an *inclusive* or.
251 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
252 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.Zero)) {
254 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
256 return InsertNewInstWith(Or, *I);
259 // If all of the demanded bits on one side are known, and all of the set
260 // bits on that side are also known to be set on the other side, turn this
261 // into an AND, as we know the bits will be cleared.
262 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
263 if (DemandedMask.isSubsetOf(RHSKnown.Zero|RHSKnown.One) &&
264 RHSKnown.One.isSubsetOf(LHSKnown.One)) {
265 Constant *AndC = Constant::getIntegerValue(VTy,
266 ~RHSKnown.One & DemandedMask);
267 Instruction *And = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
268 return InsertNewInstWith(And, *I);
271 // If the RHS is a constant, see if we can simplify it.
272 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
273 if (ShrinkDemandedConstant(I, 1, DemandedMask))
276 // If our LHS is an 'and' and if it has one use, and if any of the bits we
277 // are flipping are known to be set, then the xor is just resetting those
278 // bits to zero. We can just knock out bits from the 'and' and the 'xor',
279 // simplifying both of them.
280 if (Instruction *LHSInst = dyn_cast<Instruction>(I->getOperand(0)))
281 if (LHSInst->getOpcode() == Instruction::And && LHSInst->hasOneUse() &&
282 isa<ConstantInt>(I->getOperand(1)) &&
283 isa<ConstantInt>(LHSInst->getOperand(1)) &&
284 (LHSKnown.One & RHSKnown.One & DemandedMask) != 0) {
285 ConstantInt *AndRHS = cast<ConstantInt>(LHSInst->getOperand(1));
286 ConstantInt *XorRHS = cast<ConstantInt>(I->getOperand(1));
287 APInt NewMask = ~(LHSKnown.One & RHSKnown.One & DemandedMask);
290 ConstantInt::get(I->getType(), NewMask & AndRHS->getValue());
291 Instruction *NewAnd = BinaryOperator::CreateAnd(I->getOperand(0), AndC);
292 InsertNewInstWith(NewAnd, *I);
295 ConstantInt::get(I->getType(), NewMask & XorRHS->getValue());
296 Instruction *NewXor = BinaryOperator::CreateXor(NewAnd, XorC);
297 return InsertNewInstWith(NewXor, *I);
300 // Output known-0 bits are known if clear or set in both the LHS & RHS.
301 Known.Zero = std::move(IKnownZero);
302 // Output known-1 are known to be set if set in only one of the LHS, RHS.
303 Known.One = std::move(IKnownOne);
306 case Instruction::Select:
307 // If this is a select as part of a min/max pattern, don't simplify any
308 // further in case we break the structure.
310 if (matchSelectPattern(I, LHS, RHS).Flavor != SPF_UNKNOWN)
313 if (SimplifyDemandedBits(I, 2, DemandedMask, RHSKnown, Depth + 1) ||
314 SimplifyDemandedBits(I, 1, DemandedMask, LHSKnown, Depth + 1))
316 assert(!(RHSKnown.Zero & RHSKnown.One) && "Bits known to be one AND zero?");
317 assert(!(LHSKnown.Zero & LHSKnown.One) && "Bits known to be one AND zero?");
319 // If the operands are constants, see if we can simplify them.
320 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
321 ShrinkDemandedConstant(I, 2, DemandedMask))
324 // Only known if known in both the LHS and RHS.
325 Known.One = RHSKnown.One & LHSKnown.One;
326 Known.Zero = RHSKnown.Zero & LHSKnown.Zero;
328 case Instruction::Trunc: {
329 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
330 DemandedMask = DemandedMask.zext(truncBf);
331 Known = Known.zext(truncBf);
332 if (SimplifyDemandedBits(I, 0, DemandedMask, Known, Depth + 1))
334 DemandedMask = DemandedMask.trunc(BitWidth);
335 Known = Known.trunc(BitWidth);
336 assert(!(Known.Zero & Known.One) && "Bits known to be one AND zero?");
339 case Instruction::BitCast:
340 if (!I->getOperand(0)->getType()->isIntOrIntVectorTy())
341 return nullptr; // vector->int or fp->int?
343 if (VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
344 if (VectorType *SrcVTy =
345 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
346 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
347 // Don't touch a bitcast between vectors of different element counts.
350 // Don't touch a scalar-to-vector bitcast.
352 } else if (I->getOperand(0)->getType()->isVectorTy())
353 // Don't touch a vector-to-scalar bitcast.
356 if (SimplifyDemandedBits(I, 0, DemandedMask, Known, Depth + 1))
358 assert(!(Known.Zero & Known.One) && "Bits known to be one AND zero?");
360 case Instruction::ZExt: {
361 // Compute the bits in the result that are not present in the input.
362 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
364 DemandedMask = DemandedMask.trunc(SrcBitWidth);
365 Known = Known.trunc(SrcBitWidth);
366 if (SimplifyDemandedBits(I, 0, DemandedMask, Known, Depth + 1))
368 DemandedMask = DemandedMask.zext(BitWidth);
369 Known = Known.zext(BitWidth);
370 assert(!(Known.Zero & Known.One) && "Bits known to be one AND zero?");
371 // The top bits are known to be zero.
372 Known.Zero.setBitsFrom(SrcBitWidth);
375 case Instruction::SExt: {
376 // Compute the bits in the result that are not present in the input.
377 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
379 APInt InputDemandedBits = DemandedMask &
380 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
382 APInt NewBits(APInt::getBitsSetFrom(BitWidth, SrcBitWidth));
383 // If any of the sign extended bits are demanded, we know that the sign
385 if ((NewBits & DemandedMask) != 0)
386 InputDemandedBits.setBit(SrcBitWidth-1);
388 InputDemandedBits = InputDemandedBits.trunc(SrcBitWidth);
389 Known = Known.trunc(SrcBitWidth);
390 if (SimplifyDemandedBits(I, 0, InputDemandedBits, Known, Depth + 1))
392 InputDemandedBits = InputDemandedBits.zext(BitWidth);
393 Known = Known.zext(BitWidth);
394 assert(!(Known.Zero & Known.One) && "Bits known to be one AND zero?");
396 // If the sign bit of the input is known set or clear, then we know the
397 // top bits of the result.
399 // If the input sign bit is known zero, or if the NewBits are not demanded
400 // convert this into a zero extension.
401 if (Known.Zero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
402 // Convert to ZExt cast
403 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
404 return InsertNewInstWith(NewCast, *I);
405 } else if (Known.One[SrcBitWidth-1]) { // Input sign bit known set
406 Known.One |= NewBits;
410 case Instruction::Add:
411 case Instruction::Sub: {
412 /// If the high-bits of an ADD/SUB are not demanded, then we do not care
413 /// about the high bits of the operands.
414 unsigned NLZ = DemandedMask.countLeadingZeros();
416 // Right fill the mask of bits for this ADD/SUB to demand the most
417 // significant bit and all those below it.
418 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
419 if (ShrinkDemandedConstant(I, 0, DemandedFromOps) ||
420 SimplifyDemandedBits(I, 0, DemandedFromOps, LHSKnown, Depth + 1) ||
421 ShrinkDemandedConstant(I, 1, DemandedFromOps) ||
422 SimplifyDemandedBits(I, 1, DemandedFromOps, RHSKnown, Depth + 1)) {
423 // Disable the nsw and nuw flags here: We can no longer guarantee that
424 // we won't wrap after simplification. Removing the nsw/nuw flags is
425 // legal here because the top bit is not demanded.
426 BinaryOperator &BinOP = *cast<BinaryOperator>(I);
427 BinOP.setHasNoSignedWrap(false);
428 BinOP.setHasNoUnsignedWrap(false);
432 // If we are known to be adding/subtracting zeros to every bit below
433 // the highest demanded bit, we just return the other side.
434 if (DemandedFromOps.isSubsetOf(RHSKnown.Zero))
435 return I->getOperand(0);
436 // We can't do this with the LHS for subtraction.
437 if (I->getOpcode() == Instruction::Add &&
438 DemandedFromOps.isSubsetOf(LHSKnown.Zero))
439 return I->getOperand(1);
442 // Otherwise just hand the add/sub off to computeKnownBits to fill in
443 // the known zeros and ones.
444 computeKnownBits(V, Known, Depth, CxtI);
447 case Instruction::Shl: {
449 if (match(I->getOperand(1), m_APInt(SA))) {
451 if (match(I->getOperand(0), m_Shr(m_Value(), m_APInt(ShrAmt)))) {
452 Instruction *Shr = cast<Instruction>(I->getOperand(0));
453 if (Value *R = simplifyShrShlDemandedBits(
454 Shr, *ShrAmt, I, *SA, DemandedMask, Known))
458 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
459 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
461 // If the shift is NUW/NSW, then it does demand the high bits.
462 ShlOperator *IOp = cast<ShlOperator>(I);
463 if (IOp->hasNoSignedWrap())
464 DemandedMaskIn.setHighBits(ShiftAmt+1);
465 else if (IOp->hasNoUnsignedWrap())
466 DemandedMaskIn.setHighBits(ShiftAmt);
468 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
470 assert(!(Known.Zero & Known.One) && "Bits known to be one AND zero?");
471 Known.Zero <<= ShiftAmt;
472 Known.One <<= ShiftAmt;
473 // low bits known zero.
475 Known.Zero.setLowBits(ShiftAmt);
479 case Instruction::LShr: {
481 if (match(I->getOperand(1), m_APInt(SA))) {
482 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
484 // Unsigned shift right.
485 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
487 // If the shift is exact, then it does demand the low bits (and knows that
489 if (cast<LShrOperator>(I)->isExact())
490 DemandedMaskIn.setLowBits(ShiftAmt);
492 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
494 assert(!(Known.Zero & Known.One) && "Bits known to be one AND zero?");
495 Known.Zero.lshrInPlace(ShiftAmt);
496 Known.One.lshrInPlace(ShiftAmt);
498 Known.Zero.setHighBits(ShiftAmt); // high bits known zero.
502 case Instruction::AShr: {
503 // If this is an arithmetic shift right and only the low-bit is set, we can
504 // always convert this into a logical shr, even if the shift amount is
505 // variable. The low bit of the shift cannot be an input sign bit unless
506 // the shift amount is >= the size of the datatype, which is undefined.
507 if (DemandedMask == 1) {
508 // Perform the logical shift right.
509 Instruction *NewVal = BinaryOperator::CreateLShr(
510 I->getOperand(0), I->getOperand(1), I->getName());
511 return InsertNewInstWith(NewVal, *I);
514 // If the sign bit is the only bit demanded by this ashr, then there is no
515 // need to do it, the shift doesn't change the high bit.
516 if (DemandedMask.isSignMask())
517 return I->getOperand(0);
520 if (match(I->getOperand(1), m_APInt(SA))) {
521 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth-1);
523 // Signed shift right.
524 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
525 // If any of the high bits are demanded, we should set the sign bit as
527 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
528 DemandedMaskIn.setSignBit();
530 // If the shift is exact, then it does demand the low bits (and knows that
532 if (cast<AShrOperator>(I)->isExact())
533 DemandedMaskIn.setLowBits(ShiftAmt);
535 if (SimplifyDemandedBits(I, 0, DemandedMaskIn, Known, Depth + 1))
538 assert(!(Known.Zero & Known.One) && "Bits known to be one AND zero?");
539 // Compute the new bits that are at the top now.
540 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
541 Known.Zero.lshrInPlace(ShiftAmt);
542 Known.One.lshrInPlace(ShiftAmt);
544 // Handle the sign bits.
545 APInt SignMask(APInt::getSignMask(BitWidth));
546 // Adjust to where it is now in the mask.
547 SignMask.lshrInPlace(ShiftAmt);
549 // If the input sign bit is known to be zero, or if none of the top bits
550 // are demanded, turn this into an unsigned shift right.
551 if (BitWidth <= ShiftAmt || Known.Zero[BitWidth-ShiftAmt-1] ||
552 !DemandedMask.intersects(HighBits)) {
553 BinaryOperator *LShr = BinaryOperator::CreateLShr(I->getOperand(0),
555 LShr->setIsExact(cast<BinaryOperator>(I)->isExact());
556 return InsertNewInstWith(LShr, *I);
557 } else if (Known.One.intersects(SignMask)) { // New bits are known one.
558 Known.One |= HighBits;
563 case Instruction::SRem:
564 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
565 // X % -1 demands all the bits because we don't want to introduce
566 // INT_MIN % -1 (== undef) by accident.
567 if (Rem->isAllOnesValue())
569 APInt RA = Rem->getValue().abs();
570 if (RA.isPowerOf2()) {
571 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
572 return I->getOperand(0);
574 APInt LowBits = RA - 1;
575 APInt Mask2 = LowBits | APInt::getSignMask(BitWidth);
576 if (SimplifyDemandedBits(I, 0, Mask2, LHSKnown, Depth + 1))
579 // The low bits of LHS are unchanged by the srem.
580 Known.Zero = LHSKnown.Zero & LowBits;
581 Known.One = LHSKnown.One & LowBits;
583 // If LHS is non-negative or has all low bits zero, then the upper bits
585 if (LHSKnown.isNonNegative() || LowBits.isSubsetOf(LHSKnown.Zero))
586 Known.Zero |= ~LowBits;
588 // If LHS is negative and not all low bits are zero, then the upper bits
590 if (LHSKnown.isNegative() && LowBits.intersects(LHSKnown.One))
591 Known.One |= ~LowBits;
593 assert(!(Known.Zero & Known.One) && "Bits known to be one AND zero?");
598 // The sign bit is the LHS's sign bit, except when the result of the
599 // remainder is zero.
600 if (DemandedMask.isSignBitSet()) {
601 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1, CxtI);
602 // If it's known zero, our sign bit is also zero.
603 if (LHSKnown.isNonNegative())
604 Known.makeNonNegative();
607 case Instruction::URem: {
608 KnownBits Known2(BitWidth);
609 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
610 if (SimplifyDemandedBits(I, 0, AllOnes, Known2, Depth + 1) ||
611 SimplifyDemandedBits(I, 1, AllOnes, Known2, Depth + 1))
614 unsigned Leaders = Known2.countMinLeadingZeros();
615 Known.Zero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
618 case Instruction::Call:
619 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
620 switch (II->getIntrinsicID()) {
622 case Intrinsic::bswap: {
623 // If the only bits demanded come from one byte of the bswap result,
624 // just shift the input byte into position to eliminate the bswap.
625 unsigned NLZ = DemandedMask.countLeadingZeros();
626 unsigned NTZ = DemandedMask.countTrailingZeros();
628 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
629 // we need all the bits down to bit 8. Likewise, round NLZ. If we
630 // have 14 leading zeros, round to 8.
633 // If we need exactly one byte, we can do this transformation.
634 if (BitWidth-NLZ-NTZ == 8) {
635 unsigned ResultBit = NTZ;
636 unsigned InputBit = BitWidth-NTZ-8;
638 // Replace this with either a left or right shift to get the byte into
641 if (InputBit > ResultBit)
642 NewVal = BinaryOperator::CreateLShr(II->getArgOperand(0),
643 ConstantInt::get(I->getType(), InputBit-ResultBit));
645 NewVal = BinaryOperator::CreateShl(II->getArgOperand(0),
646 ConstantInt::get(I->getType(), ResultBit-InputBit));
648 return InsertNewInstWith(NewVal, *I);
651 // TODO: Could compute known zero/one bits based on the input.
654 case Intrinsic::x86_mmx_pmovmskb:
655 case Intrinsic::x86_sse_movmsk_ps:
656 case Intrinsic::x86_sse2_movmsk_pd:
657 case Intrinsic::x86_sse2_pmovmskb_128:
658 case Intrinsic::x86_avx_movmsk_ps_256:
659 case Intrinsic::x86_avx_movmsk_pd_256:
660 case Intrinsic::x86_avx2_pmovmskb: {
661 // MOVMSK copies the vector elements' sign bits to the low bits
662 // and zeros the high bits.
664 if (II->getIntrinsicID() == Intrinsic::x86_mmx_pmovmskb) {
665 ArgWidth = 8; // Arg is x86_mmx, but treated as <8 x i8>.
667 auto Arg = II->getArgOperand(0);
668 auto ArgType = cast<VectorType>(Arg->getType());
669 ArgWidth = ArgType->getNumElements();
672 // If we don't need any of low bits then return zero,
673 // we know that DemandedMask is non-zero already.
674 APInt DemandedElts = DemandedMask.zextOrTrunc(ArgWidth);
675 if (DemandedElts == 0)
676 return ConstantInt::getNullValue(VTy);
678 // We know that the upper bits are set to zero.
679 Known.Zero.setBitsFrom(ArgWidth);
682 case Intrinsic::x86_sse42_crc32_64_64:
683 Known.Zero.setBitsFrom(32);
687 computeKnownBits(V, Known, Depth, CxtI);
691 // If the client is only demanding bits that we know, return the known
693 if (DemandedMask.isSubsetOf(Known.Zero|Known.One))
694 return Constant::getIntegerValue(VTy, Known.One);
698 /// Helper routine of SimplifyDemandedUseBits. It computes Known
699 /// bits. It also tries to handle simplifications that can be done based on
700 /// DemandedMask, but without modifying the Instruction.
701 Value *InstCombiner::SimplifyMultipleUseDemandedBits(Instruction *I,
702 const APInt &DemandedMask,
706 unsigned BitWidth = DemandedMask.getBitWidth();
707 Type *ITy = I->getType();
709 KnownBits LHSKnown(BitWidth);
710 KnownBits RHSKnown(BitWidth);
712 // Despite the fact that we can't simplify this instruction in all User's
713 // context, we can at least compute the known bits, and we can
714 // do simplifications that apply to *just* the one user if we know that
715 // this instruction has a simpler value in that context.
716 switch (I->getOpcode()) {
717 case Instruction::And: {
718 // If either the LHS or the RHS are Zero, the result is zero.
719 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
720 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1,
723 // Output known-0 are known to be clear if zero in either the LHS | RHS.
724 APInt IKnownZero = RHSKnown.Zero | LHSKnown.Zero;
725 // Output known-1 bits are only known if set in both the LHS & RHS.
726 APInt IKnownOne = RHSKnown.One & LHSKnown.One;
728 // If the client is only demanding bits that we know, return the known
730 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
731 return Constant::getIntegerValue(ITy, IKnownOne);
733 // If all of the demanded bits are known 1 on one side, return the other.
734 // These bits cannot contribute to the result of the 'and' in this
736 if (DemandedMask.isSubsetOf(LHSKnown.Zero | RHSKnown.One))
737 return I->getOperand(0);
738 if (DemandedMask.isSubsetOf(RHSKnown.Zero | LHSKnown.One))
739 return I->getOperand(1);
741 Known.Zero = std::move(IKnownZero);
742 Known.One = std::move(IKnownOne);
745 case Instruction::Or: {
746 // We can simplify (X|Y) -> X or Y in the user's context if we know that
747 // only bits from X or Y are demanded.
749 // If either the LHS or the RHS are One, the result is One.
750 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
751 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1,
754 // Output known-0 bits are only known if clear in both the LHS & RHS.
755 APInt IKnownZero = RHSKnown.Zero & LHSKnown.Zero;
756 // Output known-1 are known to be set if set in either the LHS | RHS.
757 APInt IKnownOne = RHSKnown.One | LHSKnown.One;
759 // If the client is only demanding bits that we know, return the known
761 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
762 return Constant::getIntegerValue(ITy, IKnownOne);
764 // If all of the demanded bits are known zero on one side, return the
765 // other. These bits cannot contribute to the result of the 'or' in this
767 if (DemandedMask.isSubsetOf(LHSKnown.One | RHSKnown.Zero))
768 return I->getOperand(0);
769 if (DemandedMask.isSubsetOf(RHSKnown.One | LHSKnown.Zero))
770 return I->getOperand(1);
772 Known.Zero = std::move(IKnownZero);
773 Known.One = std::move(IKnownOne);
776 case Instruction::Xor: {
777 // We can simplify (X^Y) -> X or Y in the user's context if we know that
778 // only bits from X or Y are demanded.
780 computeKnownBits(I->getOperand(1), RHSKnown, Depth + 1, CxtI);
781 computeKnownBits(I->getOperand(0), LHSKnown, Depth + 1,
784 // Output known-0 bits are known if clear or set in both the LHS & RHS.
785 APInt IKnownZero = (RHSKnown.Zero & LHSKnown.Zero) |
786 (RHSKnown.One & LHSKnown.One);
787 // Output known-1 are known to be set if set in only one of the LHS, RHS.
788 APInt IKnownOne = (RHSKnown.Zero & LHSKnown.One) |
789 (RHSKnown.One & LHSKnown.Zero);
791 // If the client is only demanding bits that we know, return the known
793 if (DemandedMask.isSubsetOf(IKnownZero|IKnownOne))
794 return Constant::getIntegerValue(ITy, IKnownOne);
796 // If all of the demanded bits are known zero on one side, return the
798 if (DemandedMask.isSubsetOf(RHSKnown.Zero))
799 return I->getOperand(0);
800 if (DemandedMask.isSubsetOf(LHSKnown.Zero))
801 return I->getOperand(1);
803 // Output known-0 bits are known if clear or set in both the LHS & RHS.
804 Known.Zero = std::move(IKnownZero);
805 // Output known-1 are known to be set if set in only one of the LHS, RHS.
806 Known.One = std::move(IKnownOne);
810 // Compute the Known bits to simplify things downstream.
811 computeKnownBits(I, Known, Depth, CxtI);
813 // If this user is only demanding bits that we know, return the known
815 if (DemandedMask.isSubsetOf(Known.Zero|Known.One))
816 return Constant::getIntegerValue(ITy, Known.One);
825 /// Helper routine of SimplifyDemandedUseBits. It tries to simplify
826 /// "E1 = (X lsr C1) << C2", where the C1 and C2 are constant, into
827 /// "E2 = X << (C2 - C1)" or "E2 = X >> (C1 - C2)", depending on the sign
830 /// Suppose E1 and E2 are generally different in bits S={bm, bm+1,
831 /// ..., bn}, without considering the specific value X is holding.
832 /// This transformation is legal iff one of following conditions is hold:
833 /// 1) All the bit in S are 0, in this case E1 == E2.
834 /// 2) We don't care those bits in S, per the input DemandedMask.
835 /// 3) Combination of 1) and 2). Some bits in S are 0, and we don't care the
838 /// Currently we only test condition 2).
840 /// As with SimplifyDemandedUseBits, it returns NULL if the simplification was
843 InstCombiner::simplifyShrShlDemandedBits(Instruction *Shr, const APInt &ShrOp1,
844 Instruction *Shl, const APInt &ShlOp1,
845 const APInt &DemandedMask,
847 if (!ShlOp1 || !ShrOp1)
848 return nullptr; // No-op.
850 Value *VarX = Shr->getOperand(0);
851 Type *Ty = VarX->getType();
852 unsigned BitWidth = Ty->getScalarSizeInBits();
853 if (ShlOp1.uge(BitWidth) || ShrOp1.uge(BitWidth))
854 return nullptr; // Undef.
856 unsigned ShlAmt = ShlOp1.getZExtValue();
857 unsigned ShrAmt = ShrOp1.getZExtValue();
859 Known.One.clearAllBits();
860 Known.Zero.setLowBits(ShlAmt - 1);
861 Known.Zero &= DemandedMask;
863 APInt BitMask1(APInt::getAllOnesValue(BitWidth));
864 APInt BitMask2(APInt::getAllOnesValue(BitWidth));
866 bool isLshr = (Shr->getOpcode() == Instruction::LShr);
867 BitMask1 = isLshr ? (BitMask1.lshr(ShrAmt) << ShlAmt) :
868 (BitMask1.ashr(ShrAmt) << ShlAmt);
870 if (ShrAmt <= ShlAmt) {
871 BitMask2 <<= (ShlAmt - ShrAmt);
873 BitMask2 = isLshr ? BitMask2.lshr(ShrAmt - ShlAmt):
874 BitMask2.ashr(ShrAmt - ShlAmt);
877 // Check if condition-2 (see the comment to this function) is satified.
878 if ((BitMask1 & DemandedMask) == (BitMask2 & DemandedMask)) {
879 if (ShrAmt == ShlAmt)
882 if (!Shr->hasOneUse())
886 if (ShrAmt < ShlAmt) {
887 Constant *Amt = ConstantInt::get(VarX->getType(), ShlAmt - ShrAmt);
888 New = BinaryOperator::CreateShl(VarX, Amt);
889 BinaryOperator *Orig = cast<BinaryOperator>(Shl);
890 New->setHasNoSignedWrap(Orig->hasNoSignedWrap());
891 New->setHasNoUnsignedWrap(Orig->hasNoUnsignedWrap());
893 Constant *Amt = ConstantInt::get(VarX->getType(), ShrAmt - ShlAmt);
894 New = isLshr ? BinaryOperator::CreateLShr(VarX, Amt) :
895 BinaryOperator::CreateAShr(VarX, Amt);
896 if (cast<BinaryOperator>(Shr)->isExact())
897 New->setIsExact(true);
900 return InsertNewInstWith(New, *Shl);
906 /// The specified value produces a vector with any number of elements.
907 /// DemandedElts contains the set of elements that are actually used by the
908 /// caller. This method analyzes which elements of the operand are undef and
909 /// returns that information in UndefElts.
911 /// If the information about demanded elements can be used to simplify the
912 /// operation, the operation is simplified, then the resultant value is
913 /// returned. This returns null if no change was made.
914 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
917 unsigned VWidth = V->getType()->getVectorNumElements();
918 APInt EltMask(APInt::getAllOnesValue(VWidth));
919 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
921 if (isa<UndefValue>(V)) {
922 // If the entire vector is undefined, just return this info.
927 if (DemandedElts == 0) { // If nothing is demanded, provide undef.
929 return UndefValue::get(V->getType());
934 // Handle ConstantAggregateZero, ConstantVector, ConstantDataSequential.
935 if (Constant *C = dyn_cast<Constant>(V)) {
936 // Check if this is identity. If so, return 0 since we are not simplifying
938 if (DemandedElts.isAllOnesValue())
941 Type *EltTy = cast<VectorType>(V->getType())->getElementType();
942 Constant *Undef = UndefValue::get(EltTy);
944 SmallVector<Constant*, 16> Elts;
945 for (unsigned i = 0; i != VWidth; ++i) {
946 if (!DemandedElts[i]) { // If not demanded, set to undef.
947 Elts.push_back(Undef);
952 Constant *Elt = C->getAggregateElement(i);
953 if (!Elt) return nullptr;
955 if (isa<UndefValue>(Elt)) { // Already undef.
956 Elts.push_back(Undef);
958 } else { // Otherwise, defined.
963 // If we changed the constant, return it.
964 Constant *NewCV = ConstantVector::get(Elts);
965 return NewCV != C ? NewCV : nullptr;
968 // Limit search depth.
972 // If multiple users are using the root value, proceed with
973 // simplification conservatively assuming that all elements
975 if (!V->hasOneUse()) {
976 // Quit if we find multiple users of a non-root value though.
977 // They'll be handled when it's their turn to be visited by
978 // the main instcombine process.
980 // TODO: Just compute the UndefElts information recursively.
983 // Conservatively assume that all elements are needed.
984 DemandedElts = EltMask;
987 Instruction *I = dyn_cast<Instruction>(V);
988 if (!I) return nullptr; // Only analyze instructions.
990 bool MadeChange = false;
991 APInt UndefElts2(VWidth, 0);
992 APInt UndefElts3(VWidth, 0);
994 switch (I->getOpcode()) {
997 case Instruction::InsertElement: {
998 // If this is a variable index, we don't know which element it overwrites.
999 // demand exactly the same input as we produce.
1000 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1002 // Note that we can't propagate undef elt info, because we don't know
1003 // which elt is getting updated.
1004 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1005 UndefElts2, Depth + 1);
1006 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1010 // If this is inserting an element that isn't demanded, remove this
1012 unsigned IdxNo = Idx->getZExtValue();
1013 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1015 return I->getOperand(0);
1018 // Otherwise, the element inserted overwrites whatever was there, so the
1019 // input demanded set is simpler than the output set.
1020 APInt DemandedElts2 = DemandedElts;
1021 DemandedElts2.clearBit(IdxNo);
1022 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1023 UndefElts, Depth + 1);
1024 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1026 // The inserted element is defined.
1027 UndefElts.clearBit(IdxNo);
1030 case Instruction::ShuffleVector: {
1031 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1032 unsigned LHSVWidth =
1033 Shuffle->getOperand(0)->getType()->getVectorNumElements();
1034 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1035 for (unsigned i = 0; i < VWidth; i++) {
1036 if (DemandedElts[i]) {
1037 unsigned MaskVal = Shuffle->getMaskValue(i);
1038 if (MaskVal != -1u) {
1039 assert(MaskVal < LHSVWidth * 2 &&
1040 "shufflevector mask index out of range!");
1041 if (MaskVal < LHSVWidth)
1042 LeftDemanded.setBit(MaskVal);
1044 RightDemanded.setBit(MaskVal - LHSVWidth);
1049 APInt LHSUndefElts(LHSVWidth, 0);
1050 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1051 LHSUndefElts, Depth + 1);
1052 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1054 APInt RHSUndefElts(LHSVWidth, 0);
1055 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1056 RHSUndefElts, Depth + 1);
1057 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1059 bool NewUndefElts = false;
1060 unsigned LHSIdx = -1u, LHSValIdx = -1u;
1061 unsigned RHSIdx = -1u, RHSValIdx = -1u;
1062 bool LHSUniform = true;
1063 bool RHSUniform = true;
1064 for (unsigned i = 0; i < VWidth; i++) {
1065 unsigned MaskVal = Shuffle->getMaskValue(i);
1066 if (MaskVal == -1u) {
1067 UndefElts.setBit(i);
1068 } else if (!DemandedElts[i]) {
1069 NewUndefElts = true;
1070 UndefElts.setBit(i);
1071 } else if (MaskVal < LHSVWidth) {
1072 if (LHSUndefElts[MaskVal]) {
1073 NewUndefElts = true;
1074 UndefElts.setBit(i);
1076 LHSIdx = LHSIdx == -1u ? i : LHSVWidth;
1077 LHSValIdx = LHSValIdx == -1u ? MaskVal : LHSVWidth;
1078 LHSUniform = LHSUniform && (MaskVal == i);
1081 if (RHSUndefElts[MaskVal - LHSVWidth]) {
1082 NewUndefElts = true;
1083 UndefElts.setBit(i);
1085 RHSIdx = RHSIdx == -1u ? i : LHSVWidth;
1086 RHSValIdx = RHSValIdx == -1u ? MaskVal - LHSVWidth : LHSVWidth;
1087 RHSUniform = RHSUniform && (MaskVal - LHSVWidth == i);
1092 // Try to transform shuffle with constant vector and single element from
1093 // this constant vector to single insertelement instruction.
1094 // shufflevector V, C, <v1, v2, .., ci, .., vm> ->
1095 // insertelement V, C[ci], ci-n
1096 if (LHSVWidth == Shuffle->getType()->getNumElements()) {
1097 Value *Op = nullptr;
1098 Constant *Value = nullptr;
1101 // Find constant vector with the single element in shuffle (LHS or RHS).
1102 if (LHSIdx < LHSVWidth && RHSUniform) {
1103 if (auto *CV = dyn_cast<ConstantVector>(Shuffle->getOperand(0))) {
1104 Op = Shuffle->getOperand(1);
1105 Value = CV->getOperand(LHSValIdx);
1109 if (RHSIdx < LHSVWidth && LHSUniform) {
1110 if (auto *CV = dyn_cast<ConstantVector>(Shuffle->getOperand(1))) {
1111 Op = Shuffle->getOperand(0);
1112 Value = CV->getOperand(RHSValIdx);
1116 // Found constant vector with single element - convert to insertelement.
1118 Instruction *New = InsertElementInst::Create(
1119 Op, Value, ConstantInt::get(Type::getInt32Ty(I->getContext()), Idx),
1120 Shuffle->getName());
1121 InsertNewInstWith(New, *Shuffle);
1126 // Add additional discovered undefs.
1127 SmallVector<Constant*, 16> Elts;
1128 for (unsigned i = 0; i < VWidth; ++i) {
1130 Elts.push_back(UndefValue::get(Type::getInt32Ty(I->getContext())));
1132 Elts.push_back(ConstantInt::get(Type::getInt32Ty(I->getContext()),
1133 Shuffle->getMaskValue(i)));
1135 I->setOperand(2, ConstantVector::get(Elts));
1140 case Instruction::Select: {
1141 APInt LeftDemanded(DemandedElts), RightDemanded(DemandedElts);
1142 if (ConstantVector* CV = dyn_cast<ConstantVector>(I->getOperand(0))) {
1143 for (unsigned i = 0; i < VWidth; i++) {
1144 Constant *CElt = CV->getAggregateElement(i);
1145 // Method isNullValue always returns false when called on a
1146 // ConstantExpr. If CElt is a ConstantExpr then skip it in order to
1147 // to avoid propagating incorrect information.
1148 if (isa<ConstantExpr>(CElt))
1150 if (CElt->isNullValue())
1151 LeftDemanded.clearBit(i);
1153 RightDemanded.clearBit(i);
1157 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), LeftDemanded, UndefElts,
1159 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1161 TmpV = SimplifyDemandedVectorElts(I->getOperand(2), RightDemanded,
1162 UndefElts2, Depth + 1);
1163 if (TmpV) { I->setOperand(2, TmpV); MadeChange = true; }
1165 // Output elements are undefined if both are undefined.
1166 UndefElts &= UndefElts2;
1169 case Instruction::BitCast: {
1170 // Vector->vector casts only.
1171 VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1173 unsigned InVWidth = VTy->getNumElements();
1174 APInt InputDemandedElts(InVWidth, 0);
1175 UndefElts2 = APInt(InVWidth, 0);
1178 if (VWidth == InVWidth) {
1179 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1180 // elements as are demanded of us.
1182 InputDemandedElts = DemandedElts;
1183 } else if ((VWidth % InVWidth) == 0) {
1184 // If the number of elements in the output is a multiple of the number of
1185 // elements in the input then an input element is live if any of the
1186 // corresponding output elements are live.
1187 Ratio = VWidth / InVWidth;
1188 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1189 if (DemandedElts[OutIdx])
1190 InputDemandedElts.setBit(OutIdx / Ratio);
1191 } else if ((InVWidth % VWidth) == 0) {
1192 // If the number of elements in the input is a multiple of the number of
1193 // elements in the output then an input element is live if the
1194 // corresponding output element is live.
1195 Ratio = InVWidth / VWidth;
1196 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1197 if (DemandedElts[InIdx / Ratio])
1198 InputDemandedElts.setBit(InIdx);
1200 // Unsupported so far.
1204 // div/rem demand all inputs, because they don't want divide by zero.
1205 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1206 UndefElts2, Depth + 1);
1208 I->setOperand(0, TmpV);
1212 if (VWidth == InVWidth) {
1213 UndefElts = UndefElts2;
1214 } else if ((VWidth % InVWidth) == 0) {
1215 // If the number of elements in the output is a multiple of the number of
1216 // elements in the input then an output element is undef if the
1217 // corresponding input element is undef.
1218 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1219 if (UndefElts2[OutIdx / Ratio])
1220 UndefElts.setBit(OutIdx);
1221 } else if ((InVWidth % VWidth) == 0) {
1222 // If the number of elements in the input is a multiple of the number of
1223 // elements in the output then an output element is undef if all of the
1224 // corresponding input elements are undef.
1225 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1226 APInt SubUndef = UndefElts2.lshr(OutIdx * Ratio).zextOrTrunc(Ratio);
1227 if (SubUndef.countPopulation() == Ratio)
1228 UndefElts.setBit(OutIdx);
1231 llvm_unreachable("Unimp");
1235 case Instruction::And:
1236 case Instruction::Or:
1237 case Instruction::Xor:
1238 case Instruction::Add:
1239 case Instruction::Sub:
1240 case Instruction::Mul:
1241 // div/rem demand all inputs, because they don't want divide by zero.
1242 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts, UndefElts,
1244 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1245 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1246 UndefElts2, Depth + 1);
1247 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1249 // Output elements are undefined if both are undefined. Consider things
1250 // like undef&0. The result is known zero, not undef.
1251 UndefElts &= UndefElts2;
1253 case Instruction::FPTrunc:
1254 case Instruction::FPExt:
1255 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts, UndefElts,
1257 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1260 case Instruction::Call: {
1261 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1263 switch (II->getIntrinsicID()) {
1266 case Intrinsic::x86_xop_vfrcz_ss:
1267 case Intrinsic::x86_xop_vfrcz_sd:
1268 // The instructions for these intrinsics are speced to zero upper bits not
1269 // pass them through like other scalar intrinsics. So we shouldn't just
1270 // use Arg0 if DemandedElts[0] is clear like we do for other intrinsics.
1271 // Instead we should return a zero vector.
1272 if (!DemandedElts[0]) {
1274 return ConstantAggregateZero::get(II->getType());
1277 // Only the lower element is used.
1279 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1280 UndefElts, Depth + 1);
1281 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1283 // Only the lower element is undefined. The high elements are zero.
1284 UndefElts = UndefElts[0];
1287 // Unary scalar-as-vector operations that work column-wise.
1288 case Intrinsic::x86_sse_rcp_ss:
1289 case Intrinsic::x86_sse_rsqrt_ss:
1290 case Intrinsic::x86_sse_sqrt_ss:
1291 case Intrinsic::x86_sse2_sqrt_sd:
1292 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1293 UndefElts, Depth + 1);
1294 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1296 // If lowest element of a scalar op isn't used then use Arg0.
1297 if (!DemandedElts[0]) {
1299 return II->getArgOperand(0);
1301 // TODO: If only low elt lower SQRT to FSQRT (with rounding/exceptions
1305 // Binary scalar-as-vector operations that work column-wise. The high
1306 // elements come from operand 0. The low element is a function of both
1308 case Intrinsic::x86_sse_min_ss:
1309 case Intrinsic::x86_sse_max_ss:
1310 case Intrinsic::x86_sse_cmp_ss:
1311 case Intrinsic::x86_sse2_min_sd:
1312 case Intrinsic::x86_sse2_max_sd:
1313 case Intrinsic::x86_sse2_cmp_sd: {
1314 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1315 UndefElts, Depth + 1);
1316 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1318 // If lowest element of a scalar op isn't used then use Arg0.
1319 if (!DemandedElts[0]) {
1321 return II->getArgOperand(0);
1324 // Only lower element is used for operand 1.
1326 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1327 UndefElts2, Depth + 1);
1328 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1330 // Lower element is undefined if both lower elements are undefined.
1331 // Consider things like undef&0. The result is known zero, not undef.
1333 UndefElts.clearBit(0);
1338 // Binary scalar-as-vector operations that work column-wise. The high
1339 // elements come from operand 0 and the low element comes from operand 1.
1340 case Intrinsic::x86_sse41_round_ss:
1341 case Intrinsic::x86_sse41_round_sd: {
1342 // Don't use the low element of operand 0.
1343 APInt DemandedElts2 = DemandedElts;
1344 DemandedElts2.clearBit(0);
1345 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts2,
1346 UndefElts, Depth + 1);
1347 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1349 // If lowest element of a scalar op isn't used then use Arg0.
1350 if (!DemandedElts[0]) {
1352 return II->getArgOperand(0);
1355 // Only lower element is used for operand 1.
1357 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1358 UndefElts2, Depth + 1);
1359 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1361 // Take the high undef elements from operand 0 and take the lower element
1363 UndefElts.clearBit(0);
1364 UndefElts |= UndefElts2[0];
1368 // Three input scalar-as-vector operations that work column-wise. The high
1369 // elements come from operand 0 and the low element is a function of all
1371 case Intrinsic::x86_avx512_mask_add_ss_round:
1372 case Intrinsic::x86_avx512_mask_div_ss_round:
1373 case Intrinsic::x86_avx512_mask_mul_ss_round:
1374 case Intrinsic::x86_avx512_mask_sub_ss_round:
1375 case Intrinsic::x86_avx512_mask_max_ss_round:
1376 case Intrinsic::x86_avx512_mask_min_ss_round:
1377 case Intrinsic::x86_avx512_mask_add_sd_round:
1378 case Intrinsic::x86_avx512_mask_div_sd_round:
1379 case Intrinsic::x86_avx512_mask_mul_sd_round:
1380 case Intrinsic::x86_avx512_mask_sub_sd_round:
1381 case Intrinsic::x86_avx512_mask_max_sd_round:
1382 case Intrinsic::x86_avx512_mask_min_sd_round:
1383 case Intrinsic::x86_fma_vfmadd_ss:
1384 case Intrinsic::x86_fma_vfmsub_ss:
1385 case Intrinsic::x86_fma_vfnmadd_ss:
1386 case Intrinsic::x86_fma_vfnmsub_ss:
1387 case Intrinsic::x86_fma_vfmadd_sd:
1388 case Intrinsic::x86_fma_vfmsub_sd:
1389 case Intrinsic::x86_fma_vfnmadd_sd:
1390 case Intrinsic::x86_fma_vfnmsub_sd:
1391 case Intrinsic::x86_avx512_mask_vfmadd_ss:
1392 case Intrinsic::x86_avx512_mask_vfmadd_sd:
1393 case Intrinsic::x86_avx512_maskz_vfmadd_ss:
1394 case Intrinsic::x86_avx512_maskz_vfmadd_sd:
1395 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1396 UndefElts, Depth + 1);
1397 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1399 // If lowest element of a scalar op isn't used then use Arg0.
1400 if (!DemandedElts[0]) {
1402 return II->getArgOperand(0);
1405 // Only lower element is used for operand 1 and 2.
1407 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1408 UndefElts2, Depth + 1);
1409 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1410 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(2), DemandedElts,
1411 UndefElts3, Depth + 1);
1412 if (TmpV) { II->setArgOperand(2, TmpV); MadeChange = true; }
1414 // Lower element is undefined if all three lower elements are undefined.
1415 // Consider things like undef&0. The result is known zero, not undef.
1416 if (!UndefElts2[0] || !UndefElts3[0])
1417 UndefElts.clearBit(0);
1421 case Intrinsic::x86_avx512_mask3_vfmadd_ss:
1422 case Intrinsic::x86_avx512_mask3_vfmadd_sd:
1423 case Intrinsic::x86_avx512_mask3_vfmsub_ss:
1424 case Intrinsic::x86_avx512_mask3_vfmsub_sd:
1425 case Intrinsic::x86_avx512_mask3_vfnmsub_ss:
1426 case Intrinsic::x86_avx512_mask3_vfnmsub_sd:
1427 // These intrinsics get the passthru bits from operand 2.
1428 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(2), DemandedElts,
1429 UndefElts, Depth + 1);
1430 if (TmpV) { II->setArgOperand(2, TmpV); MadeChange = true; }
1432 // If lowest element of a scalar op isn't used then use Arg2.
1433 if (!DemandedElts[0]) {
1435 return II->getArgOperand(2);
1438 // Only lower element is used for operand 0 and 1.
1440 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(0), DemandedElts,
1441 UndefElts2, Depth + 1);
1442 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1443 TmpV = SimplifyDemandedVectorElts(II->getArgOperand(1), DemandedElts,
1444 UndefElts3, Depth + 1);
1445 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1447 // Lower element is undefined if all three lower elements are undefined.
1448 // Consider things like undef&0. The result is known zero, not undef.
1449 if (!UndefElts2[0] || !UndefElts3[0])
1450 UndefElts.clearBit(0);
1454 case Intrinsic::x86_sse2_pmulu_dq:
1455 case Intrinsic::x86_sse41_pmuldq:
1456 case Intrinsic::x86_avx2_pmul_dq:
1457 case Intrinsic::x86_avx2_pmulu_dq:
1458 case Intrinsic::x86_avx512_pmul_dq_512:
1459 case Intrinsic::x86_avx512_pmulu_dq_512: {
1460 Value *Op0 = II->getArgOperand(0);
1461 Value *Op1 = II->getArgOperand(1);
1462 unsigned InnerVWidth = Op0->getType()->getVectorNumElements();
1463 assert((VWidth * 2) == InnerVWidth && "Unexpected input size");
1465 APInt InnerDemandedElts(InnerVWidth, 0);
1466 for (unsigned i = 0; i != VWidth; ++i)
1467 if (DemandedElts[i])
1468 InnerDemandedElts.setBit(i * 2);
1470 UndefElts2 = APInt(InnerVWidth, 0);
1471 TmpV = SimplifyDemandedVectorElts(Op0, InnerDemandedElts, UndefElts2,
1473 if (TmpV) { II->setArgOperand(0, TmpV); MadeChange = true; }
1475 UndefElts3 = APInt(InnerVWidth, 0);
1476 TmpV = SimplifyDemandedVectorElts(Op1, InnerDemandedElts, UndefElts3,
1478 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1483 case Intrinsic::x86_sse2_packssdw_128:
1484 case Intrinsic::x86_sse2_packsswb_128:
1485 case Intrinsic::x86_sse2_packuswb_128:
1486 case Intrinsic::x86_sse41_packusdw:
1487 case Intrinsic::x86_avx2_packssdw:
1488 case Intrinsic::x86_avx2_packsswb:
1489 case Intrinsic::x86_avx2_packusdw:
1490 case Intrinsic::x86_avx2_packuswb:
1491 case Intrinsic::x86_avx512_packssdw_512:
1492 case Intrinsic::x86_avx512_packsswb_512:
1493 case Intrinsic::x86_avx512_packusdw_512:
1494 case Intrinsic::x86_avx512_packuswb_512: {
1495 auto *Ty0 = II->getArgOperand(0)->getType();
1496 unsigned InnerVWidth = Ty0->getVectorNumElements();
1497 assert(VWidth == (InnerVWidth * 2) && "Unexpected input size");
1499 unsigned NumLanes = Ty0->getPrimitiveSizeInBits() / 128;
1500 unsigned VWidthPerLane = VWidth / NumLanes;
1501 unsigned InnerVWidthPerLane = InnerVWidth / NumLanes;
1503 // Per lane, pack the elements of the first input and then the second.
1505 // v8i16 PACK(v4i32 X, v4i32 Y) - (X[0..3],Y[0..3])
1506 // v32i8 PACK(v16i16 X, v16i16 Y) - (X[0..7],Y[0..7]),(X[8..15],Y[8..15])
1507 for (int OpNum = 0; OpNum != 2; ++OpNum) {
1508 APInt OpDemandedElts(InnerVWidth, 0);
1509 for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
1510 unsigned LaneIdx = Lane * VWidthPerLane;
1511 for (unsigned Elt = 0; Elt != InnerVWidthPerLane; ++Elt) {
1512 unsigned Idx = LaneIdx + Elt + InnerVWidthPerLane * OpNum;
1513 if (DemandedElts[Idx])
1514 OpDemandedElts.setBit((Lane * InnerVWidthPerLane) + Elt);
1518 // Demand elements from the operand.
1519 auto *Op = II->getArgOperand(OpNum);
1520 APInt OpUndefElts(InnerVWidth, 0);
1521 TmpV = SimplifyDemandedVectorElts(Op, OpDemandedElts, OpUndefElts,
1524 II->setArgOperand(OpNum, TmpV);
1528 // Pack the operand's UNDEF elements, one lane at a time.
1529 OpUndefElts = OpUndefElts.zext(VWidth);
1530 for (unsigned Lane = 0; Lane != NumLanes; ++Lane) {
1531 APInt LaneElts = OpUndefElts.lshr(InnerVWidthPerLane * Lane);
1532 LaneElts = LaneElts.getLoBits(InnerVWidthPerLane);
1533 LaneElts <<= InnerVWidthPerLane * (2 * Lane + OpNum);
1534 UndefElts |= LaneElts;
1541 case Intrinsic::x86_ssse3_pshuf_b_128:
1542 case Intrinsic::x86_avx2_pshuf_b:
1543 case Intrinsic::x86_avx512_pshuf_b_512:
1545 case Intrinsic::x86_avx_vpermilvar_ps:
1546 case Intrinsic::x86_avx_vpermilvar_ps_256:
1547 case Intrinsic::x86_avx512_vpermilvar_ps_512:
1548 case Intrinsic::x86_avx_vpermilvar_pd:
1549 case Intrinsic::x86_avx_vpermilvar_pd_256:
1550 case Intrinsic::x86_avx512_vpermilvar_pd_512:
1552 case Intrinsic::x86_avx2_permd:
1553 case Intrinsic::x86_avx2_permps: {
1554 Value *Op1 = II->getArgOperand(1);
1555 TmpV = SimplifyDemandedVectorElts(Op1, DemandedElts, UndefElts,
1557 if (TmpV) { II->setArgOperand(1, TmpV); MadeChange = true; }
1561 // SSE4A instructions leave the upper 64-bits of the 128-bit result
1562 // in an undefined state.
1563 case Intrinsic::x86_sse4a_extrq:
1564 case Intrinsic::x86_sse4a_extrqi:
1565 case Intrinsic::x86_sse4a_insertq:
1566 case Intrinsic::x86_sse4a_insertqi:
1567 UndefElts.setHighBits(VWidth / 2);
1569 case Intrinsic::amdgcn_buffer_load:
1570 case Intrinsic::amdgcn_buffer_load_format:
1571 case Intrinsic::amdgcn_image_sample:
1572 case Intrinsic::amdgcn_image_sample_cl:
1573 case Intrinsic::amdgcn_image_sample_d:
1574 case Intrinsic::amdgcn_image_sample_d_cl:
1575 case Intrinsic::amdgcn_image_sample_l:
1576 case Intrinsic::amdgcn_image_sample_b:
1577 case Intrinsic::amdgcn_image_sample_b_cl:
1578 case Intrinsic::amdgcn_image_sample_lz:
1579 case Intrinsic::amdgcn_image_sample_cd:
1580 case Intrinsic::amdgcn_image_sample_cd_cl:
1582 case Intrinsic::amdgcn_image_sample_c:
1583 case Intrinsic::amdgcn_image_sample_c_cl:
1584 case Intrinsic::amdgcn_image_sample_c_d:
1585 case Intrinsic::amdgcn_image_sample_c_d_cl:
1586 case Intrinsic::amdgcn_image_sample_c_l:
1587 case Intrinsic::amdgcn_image_sample_c_b:
1588 case Intrinsic::amdgcn_image_sample_c_b_cl:
1589 case Intrinsic::amdgcn_image_sample_c_lz:
1590 case Intrinsic::amdgcn_image_sample_c_cd:
1591 case Intrinsic::amdgcn_image_sample_c_cd_cl:
1593 case Intrinsic::amdgcn_image_sample_o:
1594 case Intrinsic::amdgcn_image_sample_cl_o:
1595 case Intrinsic::amdgcn_image_sample_d_o:
1596 case Intrinsic::amdgcn_image_sample_d_cl_o:
1597 case Intrinsic::amdgcn_image_sample_l_o:
1598 case Intrinsic::amdgcn_image_sample_b_o:
1599 case Intrinsic::amdgcn_image_sample_b_cl_o:
1600 case Intrinsic::amdgcn_image_sample_lz_o:
1601 case Intrinsic::amdgcn_image_sample_cd_o:
1602 case Intrinsic::amdgcn_image_sample_cd_cl_o:
1604 case Intrinsic::amdgcn_image_sample_c_o:
1605 case Intrinsic::amdgcn_image_sample_c_cl_o:
1606 case Intrinsic::amdgcn_image_sample_c_d_o:
1607 case Intrinsic::amdgcn_image_sample_c_d_cl_o:
1608 case Intrinsic::amdgcn_image_sample_c_l_o:
1609 case Intrinsic::amdgcn_image_sample_c_b_o:
1610 case Intrinsic::amdgcn_image_sample_c_b_cl_o:
1611 case Intrinsic::amdgcn_image_sample_c_lz_o:
1612 case Intrinsic::amdgcn_image_sample_c_cd_o:
1613 case Intrinsic::amdgcn_image_sample_c_cd_cl_o:
1615 case Intrinsic::amdgcn_image_getlod: {
1616 if (VWidth == 1 || !DemandedElts.isMask())
1619 // TODO: Handle 3 vectors when supported in code gen.
1620 unsigned NewNumElts = PowerOf2Ceil(DemandedElts.countTrailingOnes());
1621 if (NewNumElts == VWidth)
1624 Module *M = II->getParent()->getParent()->getParent();
1625 Type *EltTy = V->getType()->getVectorElementType();
1627 Type *NewTy = (NewNumElts == 1) ? EltTy :
1628 VectorType::get(EltTy, NewNumElts);
1630 auto IID = II->getIntrinsicID();
1632 bool IsBuffer = IID == Intrinsic::amdgcn_buffer_load ||
1633 IID == Intrinsic::amdgcn_buffer_load_format;
1635 Function *NewIntrin = IsBuffer ?
1636 Intrinsic::getDeclaration(M, IID, NewTy) :
1637 // Samplers have 3 mangled types.
1638 Intrinsic::getDeclaration(M, IID,
1639 { NewTy, II->getArgOperand(0)->getType(),
1640 II->getArgOperand(1)->getType()});
1642 SmallVector<Value *, 5> Args;
1643 for (unsigned I = 0, E = II->getNumArgOperands(); I != E; ++I)
1644 Args.push_back(II->getArgOperand(I));
1646 IRBuilderBase::InsertPointGuard Guard(*Builder);
1647 Builder->SetInsertPoint(II);
1649 CallInst *NewCall = Builder->CreateCall(NewIntrin, Args);
1650 NewCall->takeName(II);
1651 NewCall->copyMetadata(*II);
1654 ConstantInt *DMask = dyn_cast<ConstantInt>(NewCall->getArgOperand(3));
1656 unsigned DMaskVal = DMask->getZExtValue() & 0xf;
1658 unsigned PopCnt = 0;
1659 unsigned NewDMask = 0;
1660 for (unsigned I = 0; I < 4; ++I) {
1661 const unsigned Bit = 1 << I;
1662 if (!!(DMaskVal & Bit)) {
1663 if (++PopCnt > NewNumElts)
1670 NewCall->setArgOperand(3, ConstantInt::get(DMask->getType(), NewDMask));
1675 if (NewNumElts == 1) {
1676 return Builder->CreateInsertElement(UndefValue::get(V->getType()),
1677 NewCall, static_cast<uint64_t>(0));
1680 SmallVector<uint32_t, 8> EltMask;
1681 for (unsigned I = 0; I < VWidth; ++I)
1682 EltMask.push_back(I);
1684 Value *Shuffle = Builder->CreateShuffleVector(
1685 NewCall, UndefValue::get(NewTy), EltMask);
1694 return MadeChange ? I : nullptr;